Privately Owned Vehicle Work Group Meeting - 2025/09/08 - Slot 1 & 2

[m-zain-khawaja]

Agenda

• Model integration with Autoware Universe and Autoware Core
• ZENOH integration
• SceneSeg transfer-learning off-road WG
• AutoSteer Network (EgoPath + EgoLanes)
• CARLA integration steering pipeline
• AutoSpeed
• Object Finder

Model integration with Autoware Universe and Autoware Core

@liuXinGangChina @mitsudome-r @m-zain-khawaja @pranavreddy23

Pranav has completed creating a unified ROS2 node to run all the AutoSeg networks and has optimized the visualization node with GPU-based rendering for semantic masks. This task is complete except for the new models which are being trained and will require integration.

ZENOH Integration

@evshary @m-zain-khawaja

CY has completed the Zenoh pipeline for SceneSeg which has been merged with main. He will begin work on creating the pipeline for other networks based on the same type of unified node scheme developed for ROS2.

SceneSeg transfer-learning off-road WG

@mhmdrz-mirdamadi

Muhammad Reza has created a detailed draft PR with all of the key pipeline elements for transferring the SceneSeg training pipeline to the off-road repo. The Off-Road semantic segmentation dataset has been parsed in the same format as SceneSeg and is ready for training.

AutoSteer Network (EgoPath + EgoLanes)

@m-zain-khawaja @TranHuuNhatHuy

Zain has done multiple experiments to achieve an optimal network architecture and has pushed changes to the main branch based on these experiments - the current architecture is achieving good results in the form of stable training and strong generalization on validation data. Tran has completed parsing of the CurveLanes dataset and is parsing the OpenLane dataset and now training the AutoSteer network on a combined dataset of OpenLane + TuSimple + CurveLanes.

CARLA integration steering pipeline

@JITERN

As part of CARLA integration with the Autoware POV stack steering pipeline, Jit has been working on creating ROS2 publishers and wrappers for the key data elements and nodes which will form the VisionPilot steering pipeline. He is currently working on his draft PR.

AutoSpeed

@pranavreddy23 @rsrikanth314 @auo355 @atanasko

YOLO 11 based model has been trained to achieve good results, proving that this concept works. The team is working on developing a custom bounding box detection architecture based on the YOLO 11 model which can be used for closest-in-path-object detection, as well as classification and generic object detection.

Object Finder

Muhammad Ali has begun work on Object Finder which is responsible for estimating and tracking the relative distance and speed of objects from bounding box detections in Guardian.

Attendees

Slot 1:

Slot 2:

Zoom Meeting Video Recording

Video Meeting Link
Please contact the work group lead (@m-zain-khawaja) to request access to a recording of this meeting.